Understanding the Role of Weather Data for Earth Surface Forecasting using a ConvLSTM-based Model

Abstract

Climate change is perhaps the biggest single threat to humankind and the environment, as it severely impacts our terrestrial surface, home to most of the living species. Inspired by video prediction and exploiting the availability of Copernicus Sentinel-2 images, recent studies have attempted to forecast the land surface evolution as a function of past land surface evolution, elevation, and weather. Further extending this paradigm, we propose a model based on convolutional long short-term memory (ConvLSTM) that is computationally efficient (lightweight), however obtains superior results to the previous baselines. By introducing a ConvLSTM-based architecture to this problem, we can not only ingest the heterogeneous data sources (Sentinel-2 time-series, weather data, and a Digital Elevation Model (DEM)) but also explicitly condition the future predictions on the weather. Our experiments confirm the importance of weather parameters in understanding the land cover dynamics and show that weather maps are significantly more important than the DEM in this task. Furthermore, we perform generative simulations to investigate how varying a single weather parameter can alter the evolution of the land surface. All studies are performed using the EarthNet2021 dataset. The code, additional materials and results can be found at https://github.com/dcodrut/weather2land.